DALL·E goes again

“An astronaut riding a horse.” © OpenAI.

OpenAI released a new GPT-3 version of DALL·E designed to produce images based on text descriptors. For the record, GPT-3 (Generative Pre-trained Transformer 3) is an autoregressive language model that uses deep learning to produce texts that appear to have a human author. DALL·E, based on GPT-3, is instead trained to generate pictures from descriptive phrases. You could ask it to draw a sheep or more involved, imaginary subjects such as “a tutu-wearing baby-radish walking a dog.”

“A rabbit detective sitting on a park bench and reading a newspaper in a Victorian setting.” © OpenAI, via Sam Altman.

DALL·E 2 is a higher-resolution and lower-latency version of the original. Image editing is one of its new functionalities. Starting from a photo, users select an area and ask the system to modify it. You can, for example, take a picture of a room and add a piece of furniture or mask a painting on the wall and replace it with another. As with OpenAI’s previous projects, the tool is not directly available to the public, but researchers can sign up online to get a preview of the system. OpenAI hopes to eventually make it available for use in third-party apps. Click here to find many, and often very funny, DALL·E 2 inventions.

⇨ YouTube, “DALL·E 2 Explained”

⇨ The Verge, Adi Robertson, “OpenAI’s DALL-E AI image generator can now edit pictures, too.”

2022-04-06